Confused with the models - false positives

[ghnp5]

Hello,

I'm using these model files:
https://github.com/infinitered/nsfwjs/tree/master/example/nsfw_demo/public/model
( I think it's the same as https://s3.amazonaws.com/nsfwdetector/min_nsfwjs.zip )

and I'm having several false positives.

100% non-adult face pictures get flagged as 90%+ porn

I tested also with the non-min model: https://s3.amazonaws.com/nsfwdetector/nsfwjs.zip

But, from the few examples I tried, it didn't change much at all, and some were even worse.

I see that in this issue: #276
You added some new "DAG 93" model.

Is this one supposed to have better accuracy?

What model is the best, and what should I use for best accuracy, please?

What is the difference between using the model provided in README (first link above),
and using the DAG one with { type: "graph" } ?

Are there pros and cons?

Thank you very much.

[GantMan]

The new one is supposed to have better accuracy.

Use the DAG 93. Can you provide some false positives here? I can elaborate on the pros/cons, but I want to see if it fixes your issue first.

And sorry, I think that s3 link is outdated.

[ghnp5]

Hello,
Thank you.

I've just updated to the latest version of NSFWJS, and then loaded the DAG 93 model.

Tested against a number of false positives, and the "porn" % decreased to much more acceptable values for sure.

However, I tested a few images that are 100% porn and they would give 99% on the older model, but with the DAG93 they give 72.95% for one, 56.37% for another and 34.13% for the other, which doesn't make sense. There's a 4th image that I tested that it's not porn, but a semi-naked lady, and that one gave 89.48%.

With my examples, the new model just doesn't seem accurate at all.

Happy to give you examples, please give me your email so I can send privately to you.

[GantMan]

Our latest model was provided by @TechnikEmpire so I'm tagging em here.

My email is Gant and the domain is infinite.red

[TechnikEmpire]

NSFW Post!!!

My post is NSFW because I'm going to use some real talk here to get to the nitty gritty details of what I did differently.

Alright so with the very latest model that I trained, there was a significant change.

Changes to Training Data

I used the same training data that @GantMan used in older models, but I decided to clean it up. What I did was I loaded Yahoo Open NSFW model and I deleted all images from the porn category that had a classification score of 10% or less.

I then used Yahoo model to move all images out of the sexy category into the porn folder that had a score of 60% or more.

This translated into several thousand files for each category. There were thousands of files in the sexy category that were full blown pornography. Images of sexual penetration, just plain snapshots of porn movies, full body nudity, close up of genitalia. That's not sexy, that's porn.

There were probably a couple thousand images in the porn folder of bukakke. As humans we recognize what this is, but really it's just a closeup of a face. To a neural network it's just a face, and Yahoo's Open NSFW model confirmed this by classifying them as non-pornographic. Those images were blown away, so I'm really surprised that plain faces are coming up as porn.

There were also a couple thousand images of extremely obscure pornography. I took a peek and some of them I was like "this is a completely benign image", then I'd see a 3x3 pixel area of a penis hanging out of someone's pants. This is going to throw off the neural network, at least for training purposes. Those images were blown away.

The sexy category should now only really trigger when there's provocative images. i.e. someone rolling around on a boat in a thong and bra. Nudity should come up in the pornography category.

Remarks About Changing Accuracy

With regards to the scores for categories changing, this isn't necessarily a bad thing. The way this model is to be used is that the highest scored class wins, and that's how neural networks like this are scored as well. Top-1 and Top-5 accuracy. Even if a porn image is split up like so:

Porn: 21%
Sexy: 20%
Neutral: 20%
Drawing: 20%
Hentai: 19%

The model is still accurate. It's not even necessarily indicative of a problem in the neural network.

Conclusion

What I believe happened here is twofold. First, I believe I've overfit the model by over-training it. I didn't notice this before, but your remarks made me take a second glance and it seems to be a bit overfit.

If you look at the posted Tensorflow output on this issue, you can see that train loss is decreasing while validation loss is increasing in the final train iteration.

Second, one-byte quantization would exacerbate this issue.

I think I'll run a new training session and get a new model published with one less iteration because it seems it was that last iteration that pushed the model over the edge.

Thanks for bringing this up!

[TechnikEmpire]

Oh yeah I forgot, I also started the training session with training just the final softmax layer, then fine tuning for 5 consecutive sessions, then 2 more iterations with a highly reduced LR. I'll just train it the way I did the previous model and report back.

[ghnp5]

@GantMan - just sent the email. Please let me know if you didn't receive.

Thanks @TechnikEmpire - please keep me updated!

[TechnikEmpire]

I've started training so sometime this evening I'll do a PR on the model repo.

[GantMan]

I got the email @ghnp5 thanks so much.

@TechnikEmpire - thanks for checking for overfitting! I'll look forward to your update.

[ghnp5]

Hey - any news? :)
Thanks!

[TechnikEmpire]

Yeah I've retrained, I'll check your submissions against the new model before I publish.

I also got side tracked cause I'm not happy with this 93% ceiling fine tuning so I'm training mobilenet v3 large from scratch.

[peerwaya]

Great work guys!!!

Porn: 21%
Sexy: 20%
Neutral: 20%
Drawing: 20%
Hentai: 19%

I am fairly new at this and I don't know what a good confidence score for each of the categories is.
I have used the saved_model.tflite found https://github.com/GantMan/nsfw_model/releases/tag/1.1.0 on Android/IOS with impressive results. However, I still don't know how to choose a good confidence score.

[ghnp5]

Hello,

Any news about this? :)

Thanks!

[TechnikEmpire]

@ghnp5 Yeah sorry for going MIA, I got tied up with a bunch of stuff. I have trained new models, I'm just coordinating reviewing your submission against them.